ReSS: Learning Reasoning Models for Tabular Data Prediction via Symbolic Scaffold

Dimensions overall score 8.0

• Symbolic Scaffold Chenlang Yi * 1 Gang Li * 1 Zizhan Xiong 1 Tue Minh Cao2 Yanmin Gong1 My T. Thai2 Tianbao Yang1 Abstract Tabular data remains prevalent in high-stakes do- mains such as healthcare and finance, where pre- dictive models are expected to provide both high accuracy and faithfulclaim 82%
  Evidencepartial

  Symbolic Scaffold Chenlang Yi * 1 Gang Li * 1 Zizhan Xiong 1 Tue Minh Cao2 Yanmin Gong1 My T. Thai2 Tianbao Yang1 Abstract Tabular data remains prevalent in high-stakes do- mains such as healthcare and finance

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• a sequence of tokens. 3.3. Symbolic Scaffold Informed Reasoning Dataset Curation Given a symbolic scaffold that specifies the constraints of the decision process, we leverage the input features, the output labelclaim 82%
  Evidencepartial

  a sequence of tokens. 3.3. Symbolic Scaffold Informed Reasoning Dataset Curation Given a symbolic scaffold that specifies the constraints of the decision process, we leverage the input features, the output label

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• shrink the size of training data for SFT. Below, we present an effective data augmentation strategy. A data augmentation is usually performed by perturbing the input features. However, a major challenge in data aug- mentation for tabular prediction is ensuring the consistency between perturbed features and the labelclaim 82%
  Evidencepartial

  shrink the size of training data for SFT. Below, we present an effective data augmentation strategy. A data augmentation is usually performed by perturbing the input features. However

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• Creditg 20 800 641 2564 100 100 Diabetes 8 614 536 2144 77 77 HomeLoan 11 491 406 1624 61 62 reasoning traces by setting their values to unknown in the input, and measure the resulting performance degradation. Similarly, the number of masked features is varied from k= 1 to 4claim 82%
  Evidencepartial

  Creditg 20 800 641 2564 100 100 Diabetes 8 614 536 2144 77 77 HomeLoan 11 491 406 1624 61 62 reasoning traces by setting their values to unknown in the input, and measure the resulting performance degradation. Similarly

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• mann et al., 2023) for the Diabetes and Creditg datasets. For the AD and HomeLoan datasets, we design dataset-specific serialization templates, with details provided in Appendix D. Baselines.We compare our method against a diverse set of training baselines covering classical machine learning, deep learningclaim 82%
  Evidencepartial

  mann et al., 2023) for the Diabetes and Creditg datasets. For the AD and HomeLoan datasets, we design dataset-specific serialization templates, with details provided in Appendix D. Baselines

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• base LLM, corresponding to the approach in (Xu et al., 2025). For all LLM fine-tuning methods, we use Qwen-2.5- 3B-Instruct as the base model. For RL, we use the recently proposed DisCO algorithm (Li et al., 2025), which has been observed to be better than GRPO. All hyperparameter set- tings are detailed in Appendix Cclaim 82%
  Evidencepartial

  base LLM, corresponding to the approach in (Xu et al., 2025). For all LLM fine-tuning methods, we use Qwen-2.5- 3B-Instruct as the base model. For RL, we use the recently proposed DisCO algorithm (Li et al., 2025)

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• not invent non-existent features or incorrect feature values in its reasoning. Comparison Hallucination occurs only rarely, with rates below 2% on all datasets. These results demon- strate that the generated explanations are highly faithful to the underlying input featuresclaim 82%
  Evidencepartial

  not invent non-existent features or incorrect feature values in its reasoning. Comparison Hallucination occurs only rarely, with rates below 2% on all datasets

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• generic architectures such as MLPs, which lack induc- tive biases aligned with the structure of tabular decision manifolds and often struggle to match the performance of tree-based methods (Arik & Pfister, 2020). To overcome this limitationclaim 82%
  Evidencepartial

  generic architectures such as MLPs, which lack induc- tive biases aligned with the structure of tabular decision manifolds and often struggle to match the performance of tree-based methods (Arik & Pfister, 2020)

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

Competitive landscape

ReSS bridges symbolic and neural reasoning for tabular data by using decision trees to scaffold LLMs, generating faithful and consistent natural-language explanations for high-stakes domains.